AI analysis steps
The bounded AI analysis step: what the model sees, how its citations are checked, its limits, and how its cost is capped and disclosed to buyers.
An AI analysis step asks a language model to summarise evidence your workflow has already captured. It is bounded: the model sees only the values you pass it, has no tools and no network, and every evidence id it cites is checked outside the model. This page covers when the step is available, how to configure it, what it produces and how its cost reaches buyers. It is for creators who want a written summary in their reports.
Availability#
The production server offers AI analysis. In the library, AI analysis reads "Bounded analysis of captured evidence"; on a server without a model provider it reads "Not available on this server" and cannot be added. The step's settings show the notice "Bounded AI analysis: the model sees only the captured evidence bound below, has no tools or network, and every cited evidence id and quoted value is checked outside the model."
Buyers can buy a workflow with AI analysis on SOL and USDC, where the model cost is converted at the quote's price snapshot. On ETH, a quote also needs an operator-approved model tariff for that currency; without one, the quote is refused with "The operator has not approved a model tariff for this payment asset, so AI-assisted runs cannot be quoted."
Configure the step#
Add AI analysis from the library. It depends on the most recent data or transform step and passes that step's result as a value named observed.
| Setting | Rule |
|---|---|
| Values passed to (step) | Named values from earlier steps, inputs or fixed values. Add more with New value name and Add value. "Bind tool results: only their captured evidence is shown to the model. Buyer inputs are passed as untrusted parameters, not evidence." |
| Instructions | What to write, up to 2,000 characters. New steps start with "Summarise the observed values." |
| Model | Platform default, or one of the models the server prices and allows (claude-opus-5 or claude-opus-4-8 in production). |
| Maximum output length (tokens) | 1 to 8,000. A test version needs at least 256. |
| AI cost cap per run (US$) | In Settings → Advanced limits and scope. New AI workflows start at US$0.05. |
The cost cap reads: "Every AI analysis request in a run is refused before sending once this cap would be exceeded. Paid quotes disclose it as the model maximum; production allows at most US$0.05." Set a lower cap if your step needs less.
What the model sees#
- Evidence: for each value bound to a data step (directly or through a transform), the model receives the value with the ids of the evidence it came from. Nothing else from the run is shown.
- Parameters: buyer inputs and fixed values are passed as "untrusted parameters, not evidence". The model cannot cite them.
- Your instructions, plus a note of any value that was unavailable.
The model has no tools, cannot read the chain or the web, and cannot change the workflow. The platform also caps each AI analysis step at 24,000 input tokens.
What it produces#
The model returns claims, each citing evidence ids it was given. Before anything reaches the report:
- every cited evidence id must be one supplied to this step, or the run fails with "AI analysis (step) cited evidence that was not supplied to it.";
- claims that fail the checks are withheld;
- if the model reports insufficient evidence, the step succeeds with no summary: "The model reported insufficient evidence; no summary was produced."
The step's result, Analysis result, is the accepted claims joined into one text of up to 512 characters. To show it in the report, bind it to a Text finding in a report section. The finding cites the evidence the claims cited. If no summary was produced, the finding is unknown and a required section becomes partial, so the buyer is not charged.
Every report of a workflow with AI analysis carries this limitation: "AI analysis output is bounded model synthesis of captured tool evidence. Cited evidence ids and quoted values are checked outside the model; it is not an independent observation, audit or recommendation."
Cost disclosure to buyers#
- A quote shows the model maximum as "AI-assisted summary maximum (included above)": your per-run cap converted into the payment currency, included in the execution charge cap and the maximum the buyer reserves.
- After the run, the buyer is charged the recorded model cost, rounded down to whole micro-dollars and never above the cap. The purchase form says: "AI analysis is charged at its recorded cost, up to the cap shown in the quote."
- A failed or skipped analysis, or a run that does not complete, charges nothing for the model.
- The model cost is paid to the platform; it is not part of your creator fee. See Pricing and payment rails.
Production refuses quotes above the US$0.05 cap with "This version's AI-assisted summary cap exceeds the production limit of US$0.05 per run. Create a version with a lower cap." Fix the cap and create a new version.
Review and saved examples#
- The platform checks never call a model: an AI analysis step returns one fixed fixture finding citing captured evidence, and the workflow must declare a per-run cost cap. A required finding must still come from a data step; prompt-only workflows fail.
- Workflows with AI analysis cannot publish a saved example on their listing.
The AI-assisted summary on template agents#
Token Researcher, Deployer Investigator and Wallet Analyst offer an optional summary instead of a step:
- Token Researcher: "4. Optional AI-assisted summary", with Maximum output tokens (64 to 2,000) and Maximum model cost per run · US$. The form accepts up to US$0.20, but production quotes allow at most US$0.05, so keep the cap at or below that.
- Deployer Investigator and Wallet Analyst: the checkbox "Add one capped AI-assisted summary that cites captured evidence (at most US$0.05 per run)".
The summary is one bounded model request after the reads finish. It sees only the captured evidence, makes no chain reads, and each finding must cite that evidence. Buyers see it labelled separately from the deterministic facts.